Pith. sign in

REVIEW 4 major objections 5 minor 40 references

One fully automated pipeline—CLIP pre-training, GroundingDINO-detected boxes, and modality-tagged text prompts driving MedSAM—matches or beats per-dataset specialists across 11 retinal segmentation categories.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

CLAPS unifies multi-modal retinal segmentation by auto-generating text and box prompts for SAM, and reports Dice on par with or above specialist models across 12 datasets.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A plausible automated SAM-prompting pipeline for retinal imaging, but the headline claim of matching specialist models is not supported because CLAPS trains on all 12 datasets while nnUNet trains on only one, and many gaps are under a Dice point. the 4 major comments →

arxiv 2509.08618 v1 pith:M37W3CZD submitted 2025-09-10 cs.CV

CLAPS: A CLIP-Unified Auto-Prompt Segmentation for Multi-Modal Retinal Imaging

Zhihao Zhao , Yinzheng Zhao , Junjie Yang , Xiangtong Yao , Quanmin Liang , Shahrooz Faghihroohi , Kai Huang , Nassir Navab
show 1 more author
M.Ali Nasseri
This is my paper
classification cs.CV
keywords retinal image segmentationmulti-modal imagingvision-language pre-trainingCLIPGroundingDINOSegment Anything Modelmodality ambiguityfoundation model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CLAPS is an attempt at a single, fully automated segmentation model for the retina that works across both color fundus photographs and OCT scans, covering 11 different anatomical structures and lesions. The paper claims this one pipeline matches or beats models trained separately on each dataset, and the mechanism is automation: GroundingDINO proposes lesion boxes instead of a human prompting SAM, and a learned 'modality signature' is prepended to the text prompt so a universal disease phrase is interpreted in the right imaging modality. A CLIP image-text encoder is pre-trained on 100,000 fundus and OCT images to create a shared, modality-balanced feature space before detection and segmentation are trained. If the claim holds, clinics would no longer need one model per task or modality, and SAM-based segmentation could run without manual prompting.

Core claim

On the paper's own terms, CLAPS establishes that a unified model can match or beat per-dataset specialists in retinal segmentation when three things are in place: the visual encoder is pre-trained contrastively on a large, modality-balanced retinal corpus; the SAM segmenter is prompted automatically by GroundingDINO-detected bounding boxes rather than by a human; and the text prompt is augmented with a modality signature that resolves the fact that the same clinical phrase has different visual manifestations in OCT versus fundus photography. The evaluation spans 12 public datasets and 11 segmentation categories, with CLAPS recording the best Dice score on most of them, most decisively on sma

What carries the argument

Two modules carry the argument, and one loop defines the workflow. The Modality Signature (MS) module compresses the imaging-modality label into a small dense vector, prepends it to the text token sequence, and lets self-attention fuse the two, so a single prompt such as 'possible macular hole and subretinal fluid' is read in the correct OCT or fundus context. The Adaptive Feature Fusion Module (AFFM) projects the CLIP and GroundingDINO feature maps into a common dimension, gates the CLIP channels through a sigmoid-weighted learnable mask, and lets GroundingDINO's features cross-attend to the gated CLIP features, enriching the detector's spatial proposals with CLIP's global semantics. The wo

Load-bearing premise

The central claim rests on the comparison being fair—but CLAPS trains on all 12 datasets combined, nnUNet is described as trained per dataset, and no error bars or significance tests are reported, so the small Dice gaps could reflect data access or chance rather than the architecture itself.

What would settle it

Re-train all baselines on the identical combined training set with the same augmentation and compute, repeated over several seeds, and test for significance: if unified-trained nnUNet matches or beats CLAPS, the reported gains are explained by data access, not the auto-prompt design. Separately, feed MedSAM the expert ground-truth boxes instead of GroundingDINO's detections; the resulting Dice jump measures how much of CLAPS's remaining error is detection loss, since lesions never boxed are never segmented.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single trained model can segment both fundus and OCT images across anatomy and pathology, replacing a zoo of per-task specialist models with one system.
  • SAM-based segmentation becomes fully automated: the human-in-the-loop prompt is replaced by detector boxes plus a natural-language sentence.
  • Small-lesion segmentation improves most, since box proposals locate scattered microaneurysms and fluid pockets that class-activation maps miss.
  • CLIP contrastive pre-training on a modality-balanced corpus, plus modality-weighted loss, is the ingredient the ablations credit for gains on imbalanced datasets such as IDRiD and ReTouch.
  • End-to-end training of the MedSAM stage, rather than fine-tuning a pre-trained segmenter, adds a further performance boost on imbalanced data.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The modality-signature trick is portable: any multi-modal medical setting where one disease label looks different across CT, MRI, or ultrasound could reuse the prepended-token design without changing the pipeline.
  • The real bottleneck is the detector: any lesion GroundingDINO fails to box never reaches MedSAM. Feeding ground-truth boxes instead would quantify how much of CLAPS's remaining error lives in detection rather than segmentation.
  • Because the contrastive loss ties the fused image feature to the text embedding, the model may support zero-shot segmentation of new lesion categories from a text template alone—but the paper does not test this.
  • The architecture is built to scale: as the pretraining corpus gains more modalities and datasets, the shared CLIP space should strengthen every downstream task at once—a property worth testing as the corpus grows.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. CLAPS proposes a unified, fully automated segmentation pipeline for multi-modal retinal imaging. It combines (i) a CLIP-based encoder pre-trained on a large retinal image–text corpus, (ii) a GroundingDINO detector that automatically generates bounding-box prompts from text queries, (iii) a modality-signature module that injects modality identity into text features, and (iv) MedSAM for prompt-guided segmentation. The method is evaluated on 12 public datasets spanning fundus and OCT images, reporting Dice scores against UNet, Segformer, nnUNet, and MedCLIP-SAM, with additional ablations of the main pipeline components. The central claim is that CLAPS achieves performance on par with or better than specialized expert models, thereby establishing it as a generalizable foundation model for retinal segmentation.

Significance. The paper addresses a real and important problem: automating prompt generation for SAM-based medical segmentation and unifying multiple modalities/tasks in a single framework. The proposed combination of CLIP pre-training, GroundingDINO-based spatial prompting, and a modality-signature text-embedding module is reasonable and technically well motivated. The evaluation spans a large number of datasets and lesion types, which is a valuable effort. However, the evidence for the central claim is weakened by a training-data confound between CLAPS and the specialist baselines, and by the absence of any statistical uncertainty quantification. If the authors can strengthen the evaluation, the work could make a useful contribution to automated, unified retinal segmentation.

major comments (4)
  1. [§IV-B, Tables I–IV] The head-to-head comparison confounds method with training-data volume. The paper states that nnUNet was 'trained as a specialist model on each individual dataset' while CLAPS was trained on 'the combined training sets of all datasets.' Since CLAPS sees many times more training data than any single-dataset specialist, the reported Dice differences (mostly under 1 point, e.g., ORIGA OD 96.78 vs 96.68; IDRiD HE 66.79 vs 67.51, where nnUNet is higher; ReTouch SRF 82.17 vs 83.05, where nnUNet wins) cannot be attributed to the proposed architecture, prompting strategy, or modality-signature fusion. The central claim in the abstract and §V is therefore not established by the experiments as reported.
  2. [§IV-A, §IV-B, Tables I–IV] No error bars, repeated splits, or statistical significance tests are reported for any Dice score. Many differences are sub-point, and the absence of uncertainty intervals makes it impossible to judge whether the reported improvements are meaningful or noise. The experimental settings in §IV-A describe only the 'MedCLIP model' pre-training (200 epochs, 50 warm-up, LR 5e-5) and do not document the training protocols of the baselines (optimizer, epochs, augmentation, loss weighting, compute). Without comparable tuning effort, unequal optimization cannot be excluded as an additional uncontrolled variable.
  3. [§IV-C, Tables V–VI] The ablation study varies only CLAPS's own components, so it cannot address the data-volume confound with respect to the specialist baselines. Moreover, the contribution of the Modality Signature (MS) module is not isolated: the '+P' row appears to add text prompts, but it is not stated whether MS is included or ablated. Since MS is one of the paper's stated contributions, a dedicated ablation (with vs. without MS while keeping all other components fixed) is needed to support the claim that it resolves modality ambiguity.
  4. [Abstract, §V] The paper repeatedly describes CLAPS as a 'foundation model' with 'broad generalizability,' but all experiments are within-dataset evaluations on the same datasets used for training. No zero-shot, few-shot, or cross-dataset generalization experiments are reported. The evidence supports task unification under a single set of weights, but not the stronger foundation-model generalization claim as stated.
minor comments (5)
  1. [Table I] Header says 'DIRiD' but the dataset is elsewhere referred to as 'IDRiD'. Please correct for consistency.
  2. [§IV-A] The experimental settings describe pretraining 'the MedCLIP model,' but the method uses UniMed-CLIP (§III-A) and CLIP. Clarify which model is actually used and the exact relationship between MedCLIP, UniMed-CLIP, and the CLIP encoder.
  3. [Fig. 2 and §III-D] The notation 'FF' in Fig. 2 is not defined; presumably it is the Feed-Forward Network from Eq. (9). Also, the text says the FFN compresses F_M into D_sig, but the exact dimensions of F_M and F_T are not given.
  4. [§III-A] The pretraining dataset includes 50,000 fundus images and 50,000 OCT B-scans, but the sentence says 'a large-scale and modality-balanced dataset.' It would be clearer to state explicitly how the 10:1 imbalance issue mentioned in the introduction is resolved, since the pretraining set appears balanced but the downstream datasets are not.
  5. [§I] The introduction mentions 'the typical 10:1 ratio of OCT to ICGA samples [4]' but the datasets used in this paper are fundus and OCT, not ICGA. The relevance of this example is unclear and should be either removed or properly linked to the modality imbalance issue in the evaluated data.

Circularity Check

0 steps flagged

No significant circularity: CLAPS results are supervised measurements on held-out test sets with no fitted parameter renamed as prediction and no load-bearing self-citation.

full rationale

The paper's central claims are empirical: CLAPS is trained on combined training folds and evaluated on each dataset's held-out test set. The reported Dice scores are standard supervised measurements; no parameter is fitted to the test data and then renamed a prediction. The Modality Signature (Sec. III-D) is an architectural input—a learned embedding of the modality label prepended to text features—not a quantity derived from the target segmentation; using the modality label as input does not make the output true by construction. The loss in Eq. (17) is a weighted sum of detection, segmentation, and contrastive terms, with inverse-frequency weights in Eq. (16) chosen by design rather than fitted to the evaluation. The paper contains no load-bearing self-citations: references [8], [25], and [26] are external works, and none are invoked to forbid alternatives or to define the method's conclusion. The one legitimate concern is external validity, not circularity: Sec. IV-B states nnUNet was 'trained as a specialist model on each individual dataset' while CLAPS was trained on the combined training sets, and no error bars are given, so small Dice gaps may reflect data volume or tuning effort. That is a comparison-fairness limitation and cannot be reduced to an equation equating input and output. Hence no circular step is identifiable.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on the behavior of several pretrained models and a set of unreported hyperparameters. The free parameters are mostly standard training choices, but their values are not disclosed, which harms reproducibility. The axioms reflect that the method's success depends on the transferability of CLIP and GroundingDINO to small retinal lesions, an assumption not separately validated.

free parameters (6)
  • Loss weighting hyperparameters λ_cls, λ_bbox, λ_dice, λ_bce
    Defined in Eq. 12-13 but values never reported; they are chosen by hand and affect the balance between detection and segmentation losses.
  • Contrastive temperature τ
    Appears in Eq. 14; value not stated.
  • Modality signature dimension D_sig
    Set 'to be small' but exact value not given; determines the capacity of the modality context injection.
  • Box expansion range (5%-10%) = 5%-10% random increase
    Chosen by hand to cover object edges; affects segmentation input and could impact results.
  • Number of LLM-generated text templates (10) = 10
    Chosen by hand; templates are used to create text prompts from classification labels.
  • Pre-training schedule (200 epochs, 50 warm-up, LR 5e-5) = 200 epochs, LR 5e-5
    Stated but arbitrary; affects the quality of the frozen CLIP encoder.
axioms (4)
  • domain assumption GroundingDINO can localize retinal lesions from text prompts
    The pipeline depends on the detector's ability to find small, irregular lesions (e.g., microaneurysms) using phrases like 'microaneurysms'. No evidence is given that GroundingDINO was validated for this beyond the reported results.
  • domain assumption CLIP contrastive pre-training on retinal images aligns image and text features sufficiently to improve segmentation
    The method relies on the CLIP feature space to ground both detection and segmentation; the paper assumes the pre-training transfers.
  • ad hoc to paper The LLM-generated text templates are semantically correct and varied enough to serve as training prompts
    Templates are generated by an unspecified LLM and randomly chosen; there is no analysis of their quality or coverage.
  • domain assumption The official train/test splits of the 12 datasets are comparable and combining training sets does not introduce label inconsistency
    The paper combines training sets from different datasets with potentially different annotation protocols; it assumes lesion categories like 'IRF' are labeled consistently across datasets.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CLAPS: A CLIP-Unified Auto-Prompt Segmentation for Multi-Modal Retinal Imaging." pith.science (2026). https://pith.science/paper/M37W3CZD

@misc{pith2026250908618,
  author       = {Pith},
  title        = {Pith review of: CLAPS: A CLIP-Unified Auto-Prompt Segmentation for Multi-Modal Retinal Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M37W3CZD}},
  note         = {Machine review of arXiv:2509.08618}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advancements in foundation models, such as the Segment Anything Model (SAM), have significantly impacted medical image segmentation, especially in retinal imaging, where precise segmentation is vital for diagnosis. Despite this progress, current methods face critical challenges: 1) modality ambiguity in textual disease descriptions, 2) a continued reliance on manual prompting for SAM-based workflows, and 3) a lack of a unified framework, with most methods being modality- and task-specific. To overcome these hurdles, we propose CLIP-unified Auto-Prompt Segmentation (\CLAPS), a novel method for unified segmentation across diverse tasks and modalities in retinal imaging. Our approach begins by pre-training a CLIP-based image encoder on a large, multi-modal retinal dataset to handle data scarcity and distribution imbalance. We then leverage GroundingDINO to automatically generate spatial bounding box prompts by detecting local lesions. To unify tasks and resolve ambiguity, we use text prompts enhanced with a unique "modality signature" for each imaging modality. Ultimately, these automated textual and spatial prompts guide SAM to execute precise segmentation, creating a fully automated and unified pipeline. Extensive experiments on 12 diverse datasets across 11 critical segmentation categories show that CLAPS achieves performance on par with specialized expert models while surpassing existing benchmarks across most metrics, demonstrating its broad generalizability as a foundation model.

Figures

Figures reproduced from arXiv: 2509.08618 by Junjie Yang, Kai Huang, M.Ali Nasseri, Nassir Navab, Quanmin Liang, Shahrooz Faghihroohi, Xiangtong Yao, Yinzheng Zhao, Zhihao Zhao.

Figure 1
Figure 1. Figure 1: Overview of our proposed CLAPS. Multi-modal retinal images are initially fused with feature information extracted by the CLIP encoder. Next, an Adaptive Feature Fusion Module (AFFM) fuses the global features from CLIP with the local features from GDino. Simultaneously, text and modality information are integrated through a Modality Signature (MS) module. Then, bounding box prompts are detected using the Gr… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our proposed modalities fusion enhance [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual comparative analysis of the efficacy of segmentation techniques on diverse data modalities [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 2 linked inside Pith

  1. [1]

    Segment anything in optical coherence tomography: Sam 2 for volumetric segmentation of retinal biomarkers,

    M. Kulyabin, A. Zhdanov, A. Pershin, G. Sokolov, A. Nikiforova, M. Ronkin, V . Borisov, and A. Maier, “Segment anything in optical coherence tomography: Sam 2 for volumetric segmentation of retinal biomarkers,”Bioengineering, vol. 11, no. 9, p. 940, 2024

    2024

  2. [2]

    Adapting the segment anything model for multi-modal retinal anomaly detection and localization,

    J. Li, T. Chen, X. Wang, Y . Zhong, and X. Xiao, “Adapting the segment anything model for multi-modal retinal anomaly detection and localization,”Information Fusion, vol. 113, p. 102631, 2025

    2025

  3. [3]

    A generic approach to pathological lung segmentation,

    A. Mansoor, U. Bagci, Z. Xu, B. Foster, K. N. Olivier, J. M. Elinoff, A. F. Suffredini, J. K. Udupa, and D. J. Mollura, “A generic approach to pathological lung segmentation,”IEEE transactions on medical imaging, vol. 33, no. 12, pp. 2293–2310, 2014

    2014

  4. [4]

    Oct angiography findings in acute central serous chorioretinopathy,

    N. Feucht, M. Maier, C. P. Lohmann, and L. Reznicek, “Oct angiography findings in acute central serous chorioretinopathy,”Ophthalmic Surgery, Lasers and Imaging Retina, vol. 47, no. 4, pp. 322–327, 2016

    2016

  5. [5]

    Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions,

    K. Sun, S. Xue, F. Sun, H. Sun, Y . Luo, L. Wang, S. Wang, N. Guo, L. Liu, T. Zhaoet al., “Medical multimodal foundation models in clinical diagnosis and treatment: Applications, challenges, and future directions,” arXiv preprint arXiv:2412.02621, 2024

    Pith/arXiv arXiv 2024

  6. [6]

    A comprehensive survey of foundation models in medicine,

    W. Khan, S. Leem, K. B. See, J. K. Wong, S. Zhang, and R. Fang, “A comprehensive survey of foundation models in medicine,”IEEE Reviews in Biomedical Engineering, 2025

    2025

  7. [7]

    A review on medical image segmentation: Datasets, technical models, challenges and solutions,

    H.-S. Gan, M. H. Ramlee, Z. Wang, and A. Shimizu, “A review on medical image segmentation: Datasets, technical models, challenges and solutions,”Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, vol. 15, no. 1, p. e1574, 2025

    2025

  8. [8]

    Medclip-sam: Bridging text and image towards universal medical image segmenta- tion,

    T. Koleilat, H. Asgariandehkordi, H. Rivaz, and Y . Xiao, “Medclip-sam: Bridging text and image towards universal medical image segmenta- tion,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2024, pp. 643–653

    2024

  9. [9]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,”Nature Communications, vol. 15, no. 1, p. 654, 2024

    2024

  10. [10]

    Grounding dino: Marrying dino with grounded pre-training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Suet al., “Grounding dino: Marrying dino with grounded pre-training for open-set object detection,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 38–55

    2024

  11. [11]

    Clip-art: Contrastive pre-training for fine- grained art classification,

    M. V . Conde and K. Turgutlu, “Clip-art: Contrastive pre-training for fine- grained art classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 3956–3960

    2021

  12. [12]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMedical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234–241

    2015

  13. [13]

    Unet++: A nested u-net architecture for medical image segmentation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmentation,” inInternational workshop on deep learning in medical image analysis. Springer, 2018, pp. 3–11

    2018

  14. [14]

    Atten- tion u-net: Learning where to look for the pancreas,

    O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, K. Mori, S. McDonagh, N. Y . Hammerla, B. Kainzet al., “Atten- tion u-net: Learning where to look for the pancreas,”arXiv preprint arXiv:1804.03999, 2018

    Pith/arXiv arXiv 2018

  15. [15]

    Advances in retinal microa- neurysms detection, segmentation and datasets for the diagnosis of diabetic retinopathy: a systematic literature review,

    M. Z. Tahir, M. Nasir, and S. Zhang, “Advances in retinal microa- neurysms detection, segmentation and datasets for the diagnosis of diabetic retinopathy: a systematic literature review,”Multimedia Tools and Applications, vol. 83, no. 30, pp. 74 897–74 935, 2024

    2024

  16. [16]

    Universeg: Universal medical image segmentation,

    V . I. Butoi, J. J. G. Ortiz, T. Ma, M. R. Sabuncu, J. Guttag, and A. V . Dalca, “Universeg: Universal medical image segmentation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 21 438–21 451

    2023

  17. [17]

    Unified medical image segmentation by learning from uncertainty in an end- to-end manner,

    P. Tang, P. Yang, D. Nie, X. Wu, J. Zhou, and Y . Wang, “Unified medical image segmentation by learning from uncertainty in an end- to-end manner,”Knowledge-Based Systems, vol. 241, p. 108215, 2022

    2022

  18. [18]

    Osam-fundus: A training-free, one- shot segmentation framework for optic disc and cup in fundus images,

    R. Wang, Z. Yang, and Y . Song, “Osam-fundus: A training-free, one- shot segmentation framework for optic disc and cup in fundus images,” Biomedical Signal Processing and Control, vol. 100, p. 107069, 2025

    2025

  19. [19]

    Exploring the transfer learning capabilities of clip in domain generalization for diabetic retinopathy,

    S. Baliah, F. A. Maani, S. Sanjeev, and M. H. Khan, “Exploring the transfer learning capabilities of clip in domain generalization for diabetic retinopathy,” inInternational Workshop on Machine Learning in Medical Imaging. Springer, 2023, pp. 444–453

    2023

  20. [20]

    Chatffa: interactive visual question answering on fundus fluo- rescein angiography image using chatgpt,

    X. Chen, P. Xu, Y . Li, W. Zhang, F. Song, Y .-F. Zheng, D. Shi, and M. He, “Chatffa: interactive visual question answering on fundus fluo- rescein angiography image using chatgpt,”Available at SSRN 4578568, 2023

    2023

  21. [21]

    A foundation model for generalizable disease detection from retinal images,

    Y . Zhou, M. A. Chia, S. K. Wagner, M. S. Ayhan, D. J. Williamson, R. R. Struyven, T. Liu, M. Xu, M. G. Lozano, P. Woodward-Courtet al., “A foundation model for generalizable disease detection from retinal images,”Nature, vol. 622, no. 7981, pp. 156–163, 2023

    2023

  22. [22]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4015–4026

    2023

  23. [23]

    A review of the segment anything model (sam) for medi- cal image analysis: Accomplishments and perspectives,

    M. Ali, T. Wu, H. Hu, Q. Luo, D. Xu, W. Zheng, N. Jin, C. Yang, and J. Yao, “A review of the segment anything model (sam) for medi- cal image analysis: Accomplishments and perspectives,”Computerized Medical Imaging and Graphics, vol. 119, p. 102473, 2025

    2025

  24. [24]

    Research on medical image segmentation based on sam and its future prospects,

    K. Fan, L. Liang, H. Li, W. Situ, W. Zhao, and G. Li, “Research on medical image segmentation based on sam and its future prospects,” Bioengineering, vol. 12, no. 6, p. 608, 2025

    2025

  25. [25]

    Fapn: Feature-aligned pyramid network for dense image prediction,

    S. Huang, Z. Lu, R. Cheng, and C. He, “Fapn: Feature-aligned pyramid network for dense image prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 864–873

    2021

  26. [26]

    Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,

    M. Uzair Khattak, S. Kunhimon, M. Naseer, S. Khan, and F. Shah- baz Khan, “Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities,”arXiv e-prints, pp. arXiv–2412, 2024

    2024

  27. [27]

    Origa-light: An online retinal fundus image database for glaucoma analysis and research,

    Z. Zhang, F. S. Yin, J. Liu, W. K. Wong, N. M. Tan, B. H. Lee, J. Cheng, and T. Y . Wong, “Origa-light: An online retinal fundus image database for glaucoma analysis and research,” in2010 Annual international conference of the IEEE engineering in medicine and biology. IEEE, 2010, pp. 3065–3068

    2010

  28. [28]

    Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs,

    J. I. Orlando, H. Fu, J. B. Breda, K. Van Keer, D. R. Bathula, A. Diaz- Pinto, R. Fang, P.-A. Heng, J. Kim, J. Leeet al., “Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs,”Medical image analysis, vol. 59, p. 101570, 2020

    2020

  29. [29]

    Teleophta: Machine learning and image processing methods for teleophthalmology,

    J.-C. Kleinet al., “Teleophta: Machine learning and image processing methods for teleophthalmology,”Irbm, vol. 34, no. 2, pp. 196–203, 2013

    2013

  30. [30]

    Idrid: Diabetic retinopathy–segmentation and grading challenge,

    P. Porwal, S. Pachade, M. Kokare, G. Deshmukh, J. Son, W. Bae, L. Liu, J. Wang, X. Liu, L. Gaoet al., “Idrid: Diabetic retinopathy–segmentation and grading challenge,”Medical image analysis, vol. 59, p. 101561, 2020

    2020

  31. [31]

    Diagnostic as- sessment of deep learning algorithms for diabetic retinopathy screening,

    T. Li, Y . Gao, K. Wang, S. Guo, H. Liu, and H. Kang, “Diagnostic as- sessment of deep learning algorithms for diabetic retinopathy screening,” Information Sciences, vol. 501, pp. 511–522, 2019

    2019

  32. [32]

    Learn to segment retinal lesions and beyond,

    Q. Wei, X. Li, W. Yu, X. Zhang, Y . Zhang, B. Hu, B. Mo, D. Gong, N. Chen, D. Dinget al., “Learn to segment retinal lesions and beyond,” in2020 25th International conference on pattern recognition (ICPR). IEEE, 2021, pp. 7403–7410

    2021

  33. [33]

    Retouch: The retinal oct fluid detection and segmentation benchmark and challenge,

    H. Bogunovi ´c, F. Venhuizen, S. Klimscha, S. Apostolopoulos, A. Bab- Hadiashar, U. Bagci, M. F. Beg, L. Bekalo, Q. Chen, C. Cilleret al., “Retouch: The retinal oct fluid detection and segmentation benchmark and challenge,”IEEE transactions on medical imaging, vol. 38, no. 8, pp. 1858–1874, 2019

    2019

  34. [34]

    Amd-sd: An optical coherence tomography image dataset for wet amd lesions segmentation,

    Y . Hu, Y . Gao, W. Gao, W. Luo, Z. Yang, F. Xiong, Z. Chen, Y . Lin, X. Xia, X. Yinet al., “Amd-sd: An optical coherence tomography image dataset for wet amd lesions segmentation,”Scientific Data, vol. 11, no. 1, p. 1014, 2024

    2024

  35. [35]

    Oimhs: An optical coherence tomography image dataset based on macular hole manual segmentation,

    X. Ye, S. He, X. Zhong, J. Yu, S. Yang, Y . Shen, Y . Chen, Y . Wang, X. Huang, and L. Shen, “Oimhs: An optical coherence tomography image dataset based on macular hole manual segmentation,”Scientific Data, vol. 10, no. 1, p. 769, 2023

    2023

  36. [36]

    Aroi: Annotated retinal oct images database,

    M. Melin ˇsˇcak, M. Radmilovi ˇc, Z. Vatavuk, and S. Lon ˇcari´c, “Aroi: Annotated retinal oct images database,” in2021 44th International Convention on Information, Communication and Electronic Technology (MIPRO). IEEE, 2021, pp. 371–376

    2021

  37. [37]

    Deep learning based automated detection of intraretinal cystoid fluid,

    Z. Ahmed, S. Q. Panhwar, A. Baqai, F. A. Umrani, M. Ahmed, and A. Khan, “Deep learning based automated detection of intraretinal cystoid fluid,”International Journal of Imaging Systems and Technology, vol. 32, no. 3, pp. 902–917, 2022

    2022

  38. [38]

    Dual-tree complex wavelet input transform for cyst segmentation in oct images based on a deep learning framework,

    R. Darooei, M. Nazari, R. Kafieh, and H. Rabbani, “Dual-tree complex wavelet input transform for cyst segmentation in oct images based on a deep learning framework,” inPhotonics, vol. 10, no. 1. MDPI, 2022, p. 11

    2022

  39. [39]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,”Advances in neural information processing systems, vol. 34, pp. 12 077–12 090, 2021

    2021

  40. [40]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.